Electronic Health Records (EHRs): Supporting ASCO's ... - CiteSeerX

4 downloads 10187 Views 619KB Size Report
supports data extraction and use for panomic-based clinical research, and argue for CDSS that are based on systems biology and are algorithm-based. ... personal computer's hard drive capacity (as of 2014), but a ... ble to rapid retrieval.
ELECTRONIC HEALTH RECORDS

Electronic Health Records (EHRs): Supporting ASCO’s Vision of Cancer Care Peter Yu, MD, David Artz, MD, and Jeremy Warner, MD, MS OVERVIEW ASCO’s vision for cancer care in 2030 is built on the expanding importance of panomics and big data, and envisions enabling better health for patients with cancer by the rapid transformation of systems biology knowledge into cancer care advances. This vision will be heavily dependent on the use of health information technology for computational biology and clinical decision support systems (CDSS). Computational biology will allow us to construct models of cancer biology that encompass the complexity of cancer panomics data and provide us with better understanding of the mechanisms governing cancer behavior. The Agency for Healthcare Research and Quality promotes CDSS based on clinical practice guidelines, which are knowledge bases that grow too slowly to match the rate of panomic-derived knowledge. CDSS that are based on systems biology models will be more easily adaptable to rapid advancements and translational medicine. We describe the characteristics of health data representation, a model for representing molecular data that supports data extraction and use for panomic-based clinical research, and argue for CDSS that are based on systems biology and are algorithm-based.

S

ince ASCO was founded 50 years ago, cancer care has expanded to become a global health issue to which patients, health care providers, researchers, industry, and governments have responded in an ever-increasing effort to gain control over these collective diseases of disruptive cellular growth and altered immunity. Laboratory and clinical data have generated a more comprehensive understanding of the molecular nature of malignancies—leading, we hope, to translational and personalized medicine that brings the knowledge of biologic systems to bear on the individual patient with cancer. However, these advances come at a cost and there is a risk that the benefıts will be limited to a fraction of patients with cancer or nations who can afford them. Advances in cancer treatment will need to be both sustainable and scalable. At the same time, the volume and breadth of laboratory and clinical data produced from individual patients and collectively across populations of patients, combined with the number of separate data repositories, are impediments to sharing access to data. Yet another challenge to improving cancer care outcomes is identifying and applying cancer knowledge to each individual patient when the knowledge base grows to a size that challenges human cognitive abilities. In 2013, ASCO released a document, “Shaping the Future of Oncology: Envisioning Cancer Care in 2030,” which outlines the results of the ASCO Board of Directors strategic

planning discussions.1 Big data, cancer panomics, and valuebased cancer care were identifıed as three key drivers of change that will, in large measure, shape the future of cancer care. The successful and intelligent design and implementation of health information technology (HIT) will enable these three fıelds to drive progress in cancer care. HIT is used synonymously with electronic health records (EHRs) to include all forms and uses of digital health data. EHRs are to be distinguished from electronic medical records (EMRs), which are digital patient medical records used by health care professionals and are electronic versions of a paper medical record. This article describes several developing areas of HIT that will begin to lay a foundation for the future of oncology.

WORKING WITH LARGE DATA SETS Two HIT phenomena have occurred and continue to occur, both of which are propelling medicine, and oncology in particular, rapidly into the realms of big data. The fırst is the increase in computer processing power and storage technologies roughly in line with Moore’s Law for the past 4 decades; that is, doubling about every 18 months.2,3 Although Moore’s Law did not stipulate how prices would behave in relation to transistor density, prices have decreased drastically as a function of processing power and storage. Thus, even the smallest

From the Department of Hematology and Department of Oncology, Palo Alto Medical Foundation, Mountain View, CA; Department of Medicine, Memorial Sloan Kettering Cancer Center, New York, NY; Department of Medicine and Department of Biomedical Informatics, Vanderbilt University Medical Center, Nashville, TN. Disclosures of potential conflicts of interest are found at the end of this article. Corresponding author: Peter Yu, MD, Department of Hematology and Department of Oncology, Palo Alto Medical Foundation, 701 East El Camino Real, Department of Hematology and Oncology, SV 301, Palo Alto, CA, 94301; email: [email protected]. © 2014 by American Society of Clinical Oncology.

asco.org/edbook | 2014 ASCO EDUCATIONAL BOOK

225

YU, ARTZ, AND WARNER

health care system can afford what would have seemed like limitless capabilities even 10 years prior. Second, the capacity to measure physiologic phenomena has expanded markedly over the same period. The paltry laboratory panels available in 1964 have now expanded to over 1,000 pages of laboratory tests running the gamut from acacia tree IgE to zygosity testing for multiple births, according to one recent reference.4 This does not include our expanding radiographic capabilities, immunohistochemistry, or panomic data (see below). It should be readily evident that the few scribbles of yore have been replaced with reams of digital data. The term big data is often used to describe very large and complex datasets, although there is no one point when data ceases to be merely large and becomes “big.” Rather, the definition is relative to local processing, storage, and visualization capacity. Medical data, if not truly big data, now approximates it for the individual and certainly at the population level. For example, the National Cancer Institute’s Cancer Genome Atlas (TCGA) is envisioned to eventually be 2.5 petabytes in size, about 2,500 times the size of an average personal computer’s hard drive capacity (as of 2014), but a size approachable by many modern data warehouse appliances, which are becoming increasingly integrated into cancer centers around the country and the world. To understand the implications of all of this data, structure must be disambiguated from content. We will briefly consider both of these concepts.

Data Structure All forms of computable data have structure, but the usefulness of the data can be highly variable. Colloquially, health care data is commonly referred to as structured or unstructured. Structured data is that which can be stored in an unambiguous format, usually in a data table, a list, an Extensible Markup Language (XML) fıle, or some other format amenable to rapid retrieval. Unstructured data is that which must

KEY POINTS 䡠 ASCO’s vision for cancer care emphasizes the roles of panomic data and big data. 䡠 Realizing this vision will require health information technology systems that can capture and represent panomic data in a manner that retains data meaning and allows data extraction and use. 䡠 Panomic data can be used to facilitate clinical research. 䡠 Clinical decision support systems can use panomic data to guide treatment selection and precision medicine. 䡠 The evolution of electronic health records will need to consider the rising role of panomic data in health care and research.

undergo some additional level of processing to achieve a structured format. For the purposes of this discussion, the most important unstructured data source in clinical oncology is narrative text, whether it be clinician notes, radiographic reports, pathologic reports, discharge summaries, and so forth. Add to this list of data types data which, for all practical purposes, is considered inaccessible.5 In a fully computable EMR environment, the vast majority of such data will be scanned images of text documents, usually in the Portable Document Format (PDF). Although such images may be readable to humans, they are often smudged, obscured, rasterized, skewed, or otherwise mangled to such a degree that even the most sophisticated optical character recognition system will fail to translate them (Fig. 1). Finally, a vital element of data structure is metadata, or data about data. Whether a datum is structured, unstructured, or inaccessible, the metadata about the datum (for example, the date on which it was created, the place where it was created, the categoric assignment, etc.) can be highly informative.

Data Content Structure does not defıne content, although the converse can be true. As an illustration, consider that a standard highresolution posterior-anterior and lateral chest fılm occupies about 20 MB of storage space and can be represented as a large table, where each cell is occupied by a pixel value.6 Now consider that much of this structured data comprises black pixels external to the patient, which are obviously of no importance.* Even when considering raw data, then, there may be considerable redundancies, noise, and extraneous facts. Generally, the imposition of stricter requirements for structure comes at the trade-off of expressivity. This tension is most evident in the narration of clinical encounters,7 but is applicable to newer data sources, as well. For example, TCGA uses the concept of data levels across genomic data, where the lowest level is generally raw signals and the highest level is interpreted within context (see https://tcga-data.nci.nih.gov/ tcga/tcgaDataType.jsp for more details). Content can further be understood by considering factual and contextual aspects. Consider the following three statements: • Mr. Jones has acute lymphocytic leukemia (ALL). • Mr. Jones has Philadelphia chromosome–positive (Ph⫹) ALL. • Mr. Jones has Ph⫹ ALL which is in complete remission for 1 year; he continues receiving maintenance dasatinib. The fırst statement is purely factual (ignoring the potential context of gender, for the moment). The second statement is also factual, but now brings in several contextual extenders often referred to as domain knowledge: (1) Ph⫹ ALL confers a poor prognosis, and (2) Ph⫹ ALL might have some specifıc actionable characteristics that confer sensitivity to certain drugs. The fınal statement adds additional facts but also en-

* Even this statement has its exceptions. For example, consider that someone tasked with the job of ensuring that the radiation technologists are appropriately constraining their fields of exposure may only care about the pixels external to the patient!

226

2014 ASCO EDUCATIONAL BOOK | asco.org/edbook

ELECTRONIC HEALTH RECORDS

FIG 1. An actual portion of a scanned PDF, illustrating the common phenomena of skewing and smudging. Also note the additional human annotation underlining the [20] of the karyotype result; such information may be informative but would be lost in a purely digital communication.

riches the context further; we now know that this patient beat the odds of a poor-prognosis disease, and perhaps this is the result of receiving a targeted therapy. To achieve learning from health care system data, we must have access to both facts and context. One fınal point is the importance of nonprimary source (secondary) data. Secondary data is almost always structured; the source most familiar to epidemiologists and outcomes researchers is administrative claims data (for example, International Classifıcation of Diseases, 9th Edition, Clinical Modifıcation [ICD-9-CM] codes in the United States). Although the shortcomings of these data have been welldescribed8-10, they remain easily accessible and have enabled the development of new methodologic techniques such as phenome-wide association study.11-13 Additional important sources of secondary data are population health registries. In fact, certain data elements available in tumor registries, such as the type of breast surgery performed for resectable breast cancer, are not readily available from primary data sources.** Population registries such as Surveillance, Epidemiology, and End Results (SEER) have greatly informed cancer care to date and will continue to have importance in the big-data era.

discrete data that can be used in other hospital systems to track patients with specifıc mutations (Fig. 2). Coordination between pathologists and HIT programmers has become a key element in maintaining the accuracy of the structured data so that this computer program works correctly, because the program that reads the data must be altered as changes are made to the format of the pathology report to maintain accurate data retrieval. This discrete data is then combined with the institution’s primary data storage and analysis system and is used for automated notifıcations of targeted therapy trials, usually based on automated queries of coded tumor site, coded histology, and mutations of interest. The system then generates an email notifying an investigator of new patients who meet the criteria for their study. Since molecular data management occurs centrally, study teams do not need to track and read individual pathology reports to identify patients who meet study criteria. Not only are research recruitment procedures previously dependent on manual efforts now automated, but the EHR data is liberated for other purposes such as quality assurance and clinical decision support.

Case Study: Transforming Molecular Data within an Electronic Health Record at Memorial Sloan-Kettering Cancer Center

RAPID LEARNING HEALTH SYSTEMS, CANCERLINQ™, AND CLINICAL DECISION SUPPORT

Molecular testing from mass spectrometry mutational analysis and next-generation sequencing can produce extensive lists of discrete data elements including genes, mutations, and mutational loci. By incorporating these results into the related anatomic pathology report the pathologist can disseminate a single report for both anatomic and molecular reporting to the medical oncologist, but in doing so the report is transmitted through an electronic interface to the EHR as a single document without structured data elements unless provisions are made to create this functionality. At Memorial Sloan Kettering Cancer Center (MSKCC), a computer program was created that reads the molecular report and translates the mutation data back into structured,

Knowledge results from achieving a new or improved level of understanding as a result of data analysis. Learning is the application of knowledge, in the case of health care, to improve patient care and outcomes. A learning health system is one that uses patient data routinely generated in the course of patient care delivery to generate new knowledge and learning.14 A rapid learning health system is a carefully designed ecosystem where data can flow in almost real time from generation to analysis to action. Such an ecosystem can exceed the constraints of traditional clinical trial– derived data by capturing a greater diversity of data that is meaningful to patientcentered care, such as real-world results and patient-reported outcomes.15 However, data from routine health care operations is collected in an idiosyncratic manner dependent on

** The North American Association of Central Cancer Registries (NAACCR) specifies codes that integrate the breast surgeon and reconstructive plastic surgeon’s procedures; these codes are not mappable to an individual operative report or to Current Procedural Terminology (CPT®), 4th.

asco.org/edbook | 2014 ASCO EDUCATIONAL BOOK

227

YU, ARTZ, AND WARNER

FIG 2. Data from the text pathology report is analyzed by a computer program and categorized into structured data fields. Once structured, this data can be used for future reporting and analysis. the provider’s use of syntax, the vendor and user design of EMR data capture, and the completeness and structure of data capture. ASCO believes that the future of cancer care will be reliant on a rapid learning health system for oncology that is capable of linking data sources across providers, payers, patients, government, and industry.16 CancerLinQTM is ASCO’s model for such a system and the initial design will be to improve quality of cancer care delivery, which is one of the two determinants of value-based cancer care.17 Cost of care is the other determinant in the value calculation, and rapid learning systems can secondarily provide information on measuring the cost of delivering high-quality care. Oncology clinical decision support systems (CDSS) are computerized systems that draw on knowledge bases and patient data and have the potential to tailor decisions at all phases of the cancer journey. They blend generalizable knowledge that applies to patient-defıned populations with the specifıc data linked to an individual patient.18 A wide variety of knowledge bases may be used, including peerreviewed medical literature, systematic literature reviews, knowledge based on local health care system conditions, or professional society guidance tools such as ASCO’s Provisional Clinical Opinions and Clinical Practice Guidelines. Chemotherapy selection pathways which guide treatment choices can be made computable and provide clinical decision support. The incorporation of local health care system data allows local factors that might influence health care outcomes to be included, and adds to the data available to build the rapid learning health system. Clinical practice guidelines (CPGs) are an attractive type of 228

2014 ASCO EDUCATIONAL BOOK | asco.org/edbook

knowledge base to use for CDSS. CPGs are based on systematic reviews of the literature as evaluated by an expert panel to address an important clinical decision point. Individual publications can be weighed differently based on the quality of the level of evidence and expert opinion is applied to fıll in evidentiary gaps. The Agency for Healthcare Research and Quality (AHRQ) ranks clinical practice guidelines as the highest level of generalizable knowledge base for CDSS.19 There are four limitations to most currently available CPGs: 1. Lack of transparency on the degree of reliance on expert opinion 2. Non-computable format 3. Lack of associated outcomes measures 4. Absent process for iterative guideline improvement The Institute of Medicine (IOM) has issued recommendations for writing CPGs that have the intent to ensure guideline trustworthiness and to allow evaluation of the evidence on which guideline recommendations are based. The IOM recommends that entities issuing guidelines actively monitor new literature relevant to the guideline topic and revise as appropriate, although the IOM did not recommend any formal process to actively increase the evidentiary base.20 However, in a 2012 report, the IOM was less than sanguine about the ease with which CPGs can be converted into CDSS and recommended that guideline developers address format, vocabulary, and content details in creating guidelines that will allow computer-aided clinical decision support. Despite these IOM recommendations, recent reviews of CPGs in oncology and other fıelds demonstrated that a substantial ma-

ELECTRONIC HEALTH RECORDS

jority failed to meet the IOM recommendations for trustworthy guidelines.21,22 The creation of computable CPGs begins with unambiguous statements that can be represented by clinical data that is routinely recorded in the EMR. The use of standardized vocabularies and processes to convert human language into machine-readable format can then be applied. The Guidelines into Decision Support (GLIDES) Project, in which ASCO has been a participant, is an AHRQ-sponsored project to convert CPGs into CDSS.23,24 There are four components to this model: 1. Knowledge Generation: Guideline development based on systematic literature review and expert opinion. 2. Knowledge Transformation: Use of structured vocabularies to design rules-based, machine-readable artifacts and validation by guideline developers that the artifact produced is representative of the guideline content. 3. Knowledge Integration: Build and test of the CDSS including assessment of end user factors such as usability and integration into clinical workflows. 4. Knowledge Implementation: Integration into EHRs, training of end users, and implementation into clinical use. A third problem with current CPGs is that they do not routinely incorporate outcomes measures, without which it is diffıcult to judge whether CPGs do, in fact, contribute meaningfully to clinical care. AHRQ defınes the following six types of outcomes19: 1. Clinical outcomes 2. Health care process outcomes 3. User workload and effıciency outcomes 4. Relationship-centered outcomes (patient-reported outcomes) 5. Economic outcomes 6. Use and implementation outcomes The ability to measure outcomes is a critical component for rapid learning health systems such as CancerLinQ™. Outcomes measurement allows for assessment of the variation that occurs in clinical practice concerning which guidelines are followed and when. CPGs are not meant to be inviolate or adhered to slavishly. They are reasonable recommendations that are likely to lead to desired outcomes in the clinical situation described in the guideline. However, the evidentiary base may not adequately describe the actual patient in question, or alternative clinical choices may result in equal (if not better) outcomes. Local health system resources or situations may affect best practices. The ability to measure outcomes and iteratively revise guidelines and rules-based CDSS is a singular attribute of a rapid learning health system, and the purposeful seeking out of new knowledge to fıll in knowledge gaps is fundamental to its success. There are other knowledge bases that can contribute to CancerLinQ™ and rapid learning health systems in medicine which were not identifıed by the AHRQ model. Clinical practice guideline development, even without the added steps necessary to render them machine-ready for computerized CDSS, is a laborious and time-consuming activity. By the time a CPG is issued, the recommendations may have already

become standard clinical practice, or worse, the fındings may be rendered obsolete. It has been suggested that the medical evidence for any given condition has a “half-life” of 5.5 years25; for a rapidly evolving fıeld such as oncology, the halflife is often considerably shorter. Contemporary and evidence-based guidelines may become the rate-limiting step in rapid learning health systems. Recent implementation of network meta-analytic techniques for the fırst-line treatment of chronic myelogenous leukemia (Fig. 3)26 and the adjuvant treatment of resected pancreatic adenocarcinoma27 suggest that there may be alternative approaches to guideline development grounded in the increasingly tractable quantitative analysis of primary data.

CONCLUSION We have introduced the rapidly evolving universe of oncology-specifıc EHR data, and have outlined several of ASCO’s priorities in this arena. Big data, personalized medicine, and precision oncology are no longer buzzwords, but are the current reality for a growing number of oncology practices around the country and around the world. Mechanisms which transform electronic data from the point of origin into true rapid learning health systems with integrated CDSS are the natural evolution of the DIKW stratagem, in which data is transformed into information, then into knowledge, and fınally into wisdom.28 It is essential that CDSS, a mechanism by which rapid learning health systems apply knowledge to clinical care, draw on other knowledge bases that are more nimble and are capable of driving their own growth. Cancer panomics, comprised of genetic (genomics), mRNA transcription (transcriptomics), proteins (proteomics), and metabolites (metabolomics) data, are such a knowledge base.29 The challenges of establishing lexicons and data representation and transmission standards for panomic data are signifıcant, but they are being addressed. The complexity of panomic data, and its potential to truly personalize cancer treatment, makes CDSS based on panomic data a high target value endeavor. As systems biology evolves, we will have models that allow us to better predict which therapies are most likely to succeed for a given patient. Aided by computational biology, clinically annotated panomic data will speed the development of such models. CDSS that guide appropriate ordering of molecular tests and tag results to clinical data will be a fırst step in this direction. CDSS that use highly adaptable panomic models combined with patient-level molecular data will enable a transformative step away from CDSS tied to rigid rules-based CPGs to algorithm-based CDSS that guide medical decision-making based on individual patient data.18 Because algorithm-based CDSS can be designed to consider a multitude of clinical situations, to apply patient-level data to select the clinical model that is the best fıt, and then to infer potential best choices, they begin to mimic the complex cognitive processes that oncologists work through with their patients and allow the physician to improve their situational awareness and help focus decision-making around key essentials. asco.org/edbook | 2014 ASCO EDUCATIONAL BOOK

229

YU, ARTZ, AND WARNER

FIG 3. Network analysis of 17 published regimens tested in the first-line treatment of chronic myelogenous leukemia (CML), 1968-2012. The coloration of nodes and ranking in the list on the right corresponds to the objective relative value of the regimen versus its comparators, according to a green/yellow/red schema. The coloration of edges corresponds to the strength of the outcome measure (green, overall survival; yellow, progression-free survival; red, response rate). Faded nodes and text represent regimens that have not been tested in years, implying obsolescence. See original reference for further details.

The key to creating these decision making tools starts with the curation of accurate data. Once obtained and combined with evidence-based knowledge, tremendous potential exists to analyze this data and create a rapid

learning health system. The emergence of larger data sets and greater computing power to analyze this data foresees a technology-driven driver of progress in cancer care.

Disclosures of Potential Conflicts of Interest The author(s) indicated no potential conflicts of interest.

References 1. Shaping the Future of Oncology: Envisioning Cancer Care in 2030: Outcomes of the ASCO Board of Directors Strategic Planning and Visioning Process, 2011-2012. http://www.asco.org/sites/default/fıles/ shapingfuture-lowres.pdf. Accessed March 11, 2014. 2. Moore GE. Progress in digital integrated electronics. Electron Devices Meeting, 1975 International, IEEE. 1975;21:11-13. 3. Koomey JG, Berard S, Sanchez M, et al. Implications of historical trends in the electrical effıciency of computing. Annals of the History of Computing, IEEE. 2011;33:46-54. 4. Mayo Foundation for Medical Education and Research. Rochester 2014 Interpretive Handbook. http://www.mayomedicallaboratories.com/ test-catalog/pod/MayoTestCatalog-Rochester-LaboratoryReference Edition-SortedByTestName-duplex.pdf. Accessed March 25, 2014. 5. Warner J, Hochberg E. Where is the EHR in oncology? J Natl Compr Canc Netw. 2012;10:584-588.

230

2014 ASCO EDUCATIONAL BOOK | asco.org/edbook

6. Shaw G. A clinician’s guide to digital X-ray systems. J R Soc Med. 2001; 94:391-395. 7. Rosenbloom ST, Denny JC, Xu H, et al. Data from clinical notes: a perspective on the tension between structure and flexible documentation. J Am Med Inform Assoc. 2011;18:181-186. 8. Chan M, Lim PL, Chow A, et al. Surveillance for Clostridium diffıcile infection: ICD-9 coding has poor sensitivity compared to laboratory diagnosis in hospital patients, Singapore. PloS One. 2011;6: e15603. 9. Tanpowpong P, Broder-Fingert S, Obuch JC, et al. Multicenter study on the value of ICD-9-CM codes for case identifıcation of celiac disease. Ann Epidemiol. 2013;23:136-142. 10. Liao KP, Cai T, Gainer V, et al. Electronic medical records for discovery research in rheumatoid arthritis. Arthritis Care Res (Hoboken). 2010;62: 1120-1127.

ELECTRONIC HEALTH RECORDS

11. Hebbring SJ. The challenges, advantages, and future of phenome-wide association studies. Immunology. 2014;141:157-165. 12. Warner JL, Alterovitz G. Phenome based analysis as a means for discovering context dependent clinical reference ranges. AMIA Annu Symp Proc. 2012;2012:1441-1449. 13. Denny JC, Bastarache L, Ritchie MD, et al. Systematic comparison of phenome-wide association study of electronic medical record data and genome-wide association study data. Nat Biotechnol. 2013;31:11021110. 14. Friedman CP, Wong AK, Blumenthal D. Achieving a nationwide learning health system. Sci Transl Med. 2010;2:57cm29. 15. Abernethy AP, Etheredge LM, Ganz PA, et al. Rapid-learning system for cancer care. J Clin Oncol. 2010;28:4268-4274. 16. Sledge GW, Hudis CA, Swain SM, et al. ASCO’s approach to a learning health care system in oncology. J Oncol Pract. 2013;9:145-148. 17. Sledge GW, Miller RS, Hauser R. CancerLinQ and the future of cancer care. Am Soc Clin Oncol Educ Book. 2013;430-434. 18. Eberhardt J, Bilchik A, Stojadinovic A. Clinical decision support systems: potential with pitfalls. J Surg Oncol. 2012;105:502-510. 19. Lobach D, Sanders GD, Bright TJ, et al. Enabling Health Care Decisionmaking Through Clinical Decision Support and Knowledge Management. Rockville, MD: Agency for Healthcare Research and Quality (US); 2012. 20. Graham R, Mancher M, Wolman DM, et al. Clinical Practice Guidelines We Can Trust. Washington, DC: National Academies Press; 2011.

21. Reames BN, Krell RW, Ponto SN, et al. Critical evaluation of oncology clinical practice guidelines. J Clin Oncol. 2013;31:2563-2568. 22. Kung J, Miller RR, Mackowiak PA. Failure of clinical practice guidelines to meet institute of medicine standards: Two more decades of little, if any, progress. Arch Intern Med. 2012;172:1628-1633. 23. Shiffman RN, Wright A. Evidence-based clinical decision support. Yearb Med Inform. 2013;8:120-127. 24. Yale Center for Medical Informatics. GLIDES: Guidelines Into Decision Support. http://gem.med.yale.edu/glides/. Accessed March 11, 2014. 25. Shojania KG, Sampson M, Ansari MT, et al. How quickly do systematic reviews go out of date? A survival analysis. Ann Intern Med. 2007;147: 224-233. 26. Warner J, Yang P, Alterovitz G. Automated synthesis and visualization of a chemotherapy treatment regimen network. Stud Health Technol Inform. 2013;192:62-66. 27. Liao WC, Chien KL, Lin YL, et al. Adjuvant treatments for resected pancreatic adenocarcinoma: a systematic review and network metaanalysis. Lancet Oncol. 2013;14:1095-1103. 28. Matney S, Brewster PJ, Sward KA, et al. Philosophical approaches to the nursing informatics data-information-knowledge-wisdom framework. ANS Adv Nurs Sci. 2011;34:6-18. 29. Micheel CM, Nass SJ, Omenn GS (eds). Evolution of Translational Omics: Lessons Learned and the Path Forward. Washington, DC: National Academies Press; 2012.

asco.org/edbook | 2014 ASCO EDUCATIONAL BOOK

231